Jina's remote MCP server filters its own tools to save context

What it does

A hosted Model Context Protocol server that exposes Jina AI’s Reader, Embeddings, and Reranker APIs as nineteen callable tools. Your LLM can scrape web pages to markdown, search the web or arXiv, capture screenshots, rerank documents, classify text, deduplicate images or strings, and extract figures from PDFs. A subset of tools works without an API key at reduced rate limits; the rest require a Jina key.

The interesting bit

The server can filter its own tool roster before registration via URL query parameters—exclude_tools, include_tags, and so on—so unwanted tools never reach the client and don’t consume context-window tokens. It’s a pragmatic admission that nineteen tool schemas can overwhelm smaller context windows.

Key highlights

• 19 tools covering search, reading, rerank, embeddings, and utility operations
• Optional API key for basic read and primer tools; required for search, embeddings, and reranking
• Server-side tool filtering by name or tag to reduce context-window bloat
• Streamable HTTP transport (MCP spec 2025-03-26) with an /sse alias for backward compatibility
• Parallel variants for search and read operations, though the README notes LLMs rarely use them unless explicitly prompted

Caveats

• Some MCP clients don’t support environment variables, which can force hardcoded API keys
• Client-specific quirks are documented, including Windows argument-escaping bugs and Cursor UI glitches
• LLMs rarely invoke the parallel tool variants unless explicitly instructed

Verdict

Useful if you want hosted web search, academic search, and semantic utilities inside your MCP client without self-hosting. Less compelling if you already have local search and embedding infrastructure you prefer.